Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection
Hamid Reza Firoozfar, Mohammadsadegh Abolhasani, Reza Mousavi, Paul Jen-Hwa Hu
Read on arXiv →
Key claim
New taxonomy improves detection of coded language in moderation.
This paper presents a new way to categorize indirect linguistic expressions used on social media, showing a 4.7% improvement in accuracy and 5.4% in F1 score over existing methods. This could significantly enhance content moderation systems by providing a clearer framework for understanding coded language.
In plain English
Imagine you're trying to keep track of how people communicate sensitive topics on social media without getting flagged for moderation. Users often come up with clever ways to say things indirectly, using euphemisms or coded language to mask their true intent. This can make it really hard for content moderation systems to catch harmful content because the language is so nuanced and context-dependent. When these systems rely on existing frameworks, they often miss the mark, leading to either over-moderation or under-moderation — this is what's called a failure to detect algospeak or adversarial obfuscation effectively.
To address this, the authors propose a new way to categorize these indirect expressions based on the mechanisms behind how meaning is encoded and decoded, rather than just the communicative goals. They developed a comprehensive taxonomy that helps to systematically identify and analyze these expressions. By testing this taxonomy with 2,000 posts from platforms like TikTok and Bluesky, they found that it significantly outperformed previous taxonomies, achieving better accuracy and F1 scores.
In practical terms, this means that if you're building a content moderation tool, using this new taxonomy could help you better detect and understand the emerging coded language that users employ to evade moderation. It provides a more stable framework for identifying harmful content while respecting the nuances of user expression.
The paper introduces a new taxonomy for indirect linguistic expressions, which is a meaningful extension of existing frameworks.
The evaluation against multiple existing taxonomies and a solid dataset supports the claims made.
Deep reliability assessment
The methodology supports the narrower claim that, on their 2,000 English TikTok and Bluesky posts, a comprehensive mechanism-oriented taxonomy improves prompt-based LLM detection versus four prior taxonomies and a no-taxonomy prompt. It overreaches if read as proving robust moderation performance in production, because it does not test multimodal content, non-English or code-switched settings, adversarial adaptation after deployment, or stronger non-prompt baselines.
Reproducibility
Partially reproducible: code and the publicly releasable portion of the dataset are released on GitHub, but TikTok-derived text is not redistributed because of data-sharing restrictions.
Discussion questions
- 1.The paper frames taxonomy quality as the key driver of LLM performance, but the proposed taxonomy is also the annotation scheme used to label the dataset. Is that a fair evaluation, or does it bake in an advantage over older taxonomies that were not designed around these labels?
- 2.They compare taxonomy-guided prompting against other taxonomy prompts and a no-taxonomy prompt. If you were building this for moderation, would that be the fair fight, or would you expect retrieval over fresh platform slang, human-in-the-loop review, or a fine-tuned classifier to be the real baseline?
- 3.The authors intentionally abstract away from communicative intent and classify only the mechanism of encoding. In an actual product policy workflow, is that separation useful, or does moderation need intent and harm context so badly that mechanism-only detection risks creating noisy alerts?
- 4.The evaluation is English-only and text-only, while a lot of coded language in SEA social platforms is multilingual, code-switched, image/video/audio-heavy, and locally contextual. Does the top-level mechanism taxonomy still feel portable, or would you expect the hardest cases to break it?
- 5.The best reported gains are meaningful but not huge: the abstract claims 4.7 percentage points in accuracy and 5.4 in F1 over the best benchmark, while GPT-5.4 gains over the strongest taxonomy are smaller. What deployment threshold would make this worth adding to a moderation pipeline, given the likely cost, latency, and false-positive risk?
Key figure
Figure 1 shows a TikTok example where the direct term porn star returns no results, while the same meaning is encoded through a sequence of emojis combining several mechanisms such as pictorial resemblance and name-based mapping.